ALLAN A.
NIELSEN: GEOSTATISTICS AND ANALYSIS OF SPATIAL DATA
Geostatistics and Analysis of Spatial Data
Allan A. Nielsen
Abstract— This note deals with geostatistical measures for spatial cor-
relation, namely the auto-covariance function and the semi-variogram, as
well as deterministic and geostatistical methods for spatial interpolation,
namely inverse distance weighting, radial basis functions (RBF) and krig-
ing. Some semi-variogram models are mentioned, specifically the spherical,
the exponential and the Gaussian models. Equations for RBF interpolation
as well as simple and ordinary kriging (OK) are deduced. Other types of
kriging are mentioned, and references to international literature, Internet
addresses and state-of-the-art software in the field are given. A very simple
example to illustrate the computations for OK and a more realistic exam-
ple with height data from an area near Slagelse, Denmark, are given. A
series of attractive characteristics of kriging are mentioned, and a simple
sampling strategic consideration is given based on the dependence of the
kriging variance of distance and direction to the nearest observations.
I. I NTRODUCTION
O FTEN we need to be able to integrate point attribute infor-
mation with vector and raster data which we may already
have stored in a Geographical Information System (GIS). This
can be done by linking the point information to a geographi-
cal coordinate in the data base. If we have lots of point data, a
tempting alternative will be to generate an interpolated map so
that from our point data we calculate raster data which can be
analysed along with other sources of raster data.
This note deals with geostatistical methods for description of
spatial correlation between point measurements as well as de-
terministic and geostatical methods for spatial interpolation.
The basic idea in geostatistics consists of considering ob-
served values of geochemical, geophysical or other natural vari-
ables as realisations of a stochastic process in the 2-D plane or
in 3-D space. For each position r in a domain D which is a
part of Euclidian space, a measureable quantity z(r) termed a
regionalised variable exists. z(r) is considered as a realisa-
tion of a stochastic variable Z(r). The set of stochastic vari-
ables {Z(r) | r ∈ D} is termed a stochastic function. Z(r) has
mean value or expectation value E{Z(r)} = µ(r) and auto-
covariance function Cov{Z(r), Z(r + h)} = C(r, h), where
h is termed the displacement vector. If µ(r) is constant over
D, i.e., µ(r) = µ, Z is said to be first order stationary. If also
C(r, h) is constant over D, i.e., C(r, h) = C(h), Z is said to
be second order stationary.
This statistical view is inspired by work carried out by
Georges Matheron in 1962-1963. It is described in for example
[1], [2]. [3] gives a good practical and data analytically oriented
introduction to geostatistics. [4] is a chapter in a collection of
articles which describe many different techniques and their ap-
plication within the geosciences. [5] deals with geostatistics and
other relevant subjects in the context of analysis of spatial data.
Geostatistical expositions in a GIS context can be found in [6],
[7]. [8] deals with multivariate geostatistics, i.e., studies of the
joint spatial co-variation of more variables. The International
Association for Mathematical Geology (IAMG) publishes i.a.
the periodical Mathematical Geosciences (formerly Mathemat-
ical Geology) where many results on geostatistical research are
published. State-of-the-art open source software may be found
in GSLIB, [9], and Variowin, [10]. Other easily obtainable soft-
wares are Geo-EAS and Geostatistical Toolbox. All these pack-
Allan Aasbjerg Nielsen, DTU Space – National Space Institute, Tech-
nical University of Denmark, Building 321, DK-2800 Lyngby, Telephone
+45 4525 3425, Telefax +45 4588 1397, E-mail: aa@space.dtu.dk, Internet
http://www.imm.dtu.dk/∼aa. 6 October 2009.
1
ages can be found at http://www-sst.unil.ch/research/variowin/
(or via a search engine). Also commercial geostatistical soft-
ware exists.
This note which is inspired by [11] (see also [12]), in Sec-
tion II deals with spatial correlation, specifically the auto-
covariance function, the semi-variogram and some semi-
variogram models are described. Section III deals with spatial
interpolation including the deterministic methods inverse dis-
tance weighting and radial basis function interpolation along
with a family of statistically based methods termed kriging.
Here simple and ordinary kriging are dealt with in some detail.
Section IV gives final remarks.
II. S PATIAL C ORRELATION
This section mentions methods for description of similarity
between measurements of natural variables in the 2-D plane or
in 3-D space. Specifically the auto-covariance function and the
semi-variogram are introduced. Also a relation between the two
is given.
A. The Semi-Variogram
Consider two scalar quantities z(r) and z(r + h) measured
at two points in the plane or in space r og r + h separated by
the displacement vector h. We consider z as a realisation of a
stochastic variable Z. The variability may be described b.m.o.
the auto-covariance function (assuming or enforcing first order
stationarity, i.e., the mean value is position independent)
C(r, h) = E{[Z(r) − µ][Z(r + h) − µ]}.
The variogram, 2γ, is defined as
2γ(r, h) = E{[Z(r) − Z(r + h)]2 },
which is a measure for the expected squared difference between
stochastic variables as a function of position and the displace-
ment vector. In general the variogram will depend on the dis-
placement vector h as well as on the position vector r. The in-
trinsic hypothesis of geostatistics says that the semi-variogram,
γ, is independent of the position vector and that it depends only
on the displacement vector, i.e.,
γ(r, h) = γ(h).
If Z(r) is second order stationary (i.e., its auto-covariance
function is position independent), the intrinsic hypothesis is
valid whereas the opposite is not necessarily true.
If we assume or enforce second order stationarity the follow-
ing relation between the auto-covariance function and the semi-
variogram is valid
γ(h) = C(0) − C(h). (1)
Note, that C(0) = σ 2 , the variance of the stochastic variable.
Given a set of point measurements the semi-variogram may
be calculated b.m.o. the following estimator, which calculates
(half) the mean value of the squared differences between all
pairs of measurements z(r k ) and z(r k + h) separated by the
displacement vector h
N (h)
1 X
γ̂(h) = [z(r k ) − z(r k + h)]2 .
2N (h)
k=1
2 ALLAN A. NIELSEN: GEOSTATISTICS AND ANALYSIS OF SPATIAL DATA

N (h) is the number of point pairs separated by h. γ̂ is termed Z1 Z2 Z0 Z3

the experimental semi-variogram. Often we calculate mean val-
ues of γ̂ over intervals h ± ∆h for both length (magnitude)
and direction (argument) of h. Mean values for the magni-
tude of h (h ± ∆h) are calculated in order to get a sufficiently
high N (h) to obtain a low estimation variance for the semi-
variogram value. Mean values over intervals af the argument of
h are calculated to check for possible anisotropy. Anisotropy
refers to the characteristic that the auto-covariance function and
the semi-variogram do not behave similarly for all directions
of the displacement vector between observations. This possi-
ble anisotropy may also be checked by calculating 2-D semi-
variograms also known as variogram maps, [3], [13], [11], [10],
[9].
B. Semi-Variogram Models
In order to be able to define its characteristics we parameterise
the semi-variogram b.m.o. different semi-variogram models. An
often used model, γ ∗ , is the spherical model (here we assume or
impose isotropy, i.e., the semi-variogram depends only on dis-
tance and not on direction between observations, and we denote
by h the magnitude of h)
 1 3<h≤6 2
 0
 h i h=0
∗
γ (h) = 3 h
C0 + C1 2 R − 2 R 31 h3
0<h<R

 sample sites. Each circle is centered on a sample point and its
C0 + C1 h≥R,
where C0 is the so-called nugget effect and R is termed the
range of influence or just the range; C0 /(C0 +C1 ) is the relative
nugget effect and C0 + C1 is termed the sill (= σ 2 ). The param-
eters C0 and C1 are not to be confused with the auto-covariance
function C(h). The nugget effect is a discontinuity in the semi-
variogram for h = 0, which is due to both measurement uncer-
tainties and micro variability that cannot be revealed at the scale
of sampling. The range of influence is the distance where co-
variation between samples ceases to exist; measurements taken
further apart are uncorrelated.
Two other models often used are the exponential model (see
Figure 4)
½
∗ 0 £ ¡ ¢¤ h = 0
γ (h) =
C0 + C1 1 − exp − 3h R h>0
-2 -1 0 1 2 3
Fig. 1. Simple example with three observations.
For examples on an experimental semi-variogram and differ-
ent models, see Figures 4 and 5.
Note, that C(0) is the auto-covariance function for displace-
ment vector h = 0, and that C0 is a parameter in the semi-
variogram model.
C. Examples
To illustrate the calculations Figure 1 shows a very simple
example with three observations, z1 = 1, z2 = 3 og z3 = 2
with (1-D) coordinates –2, –1 og 3. The semi-variogram γ̂ with
∆h = 1.5 is calculated like this (lags are distance groups de-
fined by h ± ∆h)
lag h N γ̂
0 0<h≤3 1 1/2(1 − 3)2 = 2
1/4((1 − 2)2 + (3 − 2)2 ) = 1/2
As another more realistic example Figure 2 shows a map with
radius is proportional to the quantity measured, in this case the
height above the ground water in a 10 km × 10 km area near
Slagelse, Denmark. Figure 3 shows a histogram for these data.
Figure 4 shows all possible pairwise squared differences as
a function of distance between observations for the height data
(assuming isotropy). Also an exponential variogram model es-
timated directly on this point cloud is shown. The nugget effect
is 0 m2 , the effective range is 3,840 m and the sill is 840 m2
(corresponding to 420 m2 for the semi-variogram model).
Figure 5 shows the corresponding experimental semi-
variogram. ∆h is here 100 m and again we assume isotropy.
Traditionally the first lag interval is half the size of the remain-
ing lags, here 50 m. The experimental semi-variogram indicates
146000

and the Gaussian model (see Figure 5)

(
∗
0 h ³ ´i h = 0
γ (h) =
144000

3h2
C0 + C1 1 − exp − R2 h>0.

These latter two models never reach but approach the sill asymp-
Northing

142000

totically. Due to its horisontal tangent for h → 0 the Gaussian

model is good for describing very continuous phenomena.
Also other semi-variogram models such as linear and power
140000

functions are some times applied. To allow for so-called nested

structures where the semi-variogram has different structures de-
pending on the magnitude and possibly the direction of the dis-
placement vector between observations, combinations of mod-
138000

els may be useful.

The model parameters may be estimated b.m.o. iterative, non-
linear least squares methods. These minimise the squared differ-
644000 646000 648000 650000 652000 654000
ences between the experimental semi-variogram and the model
considered as a function of the vector of parameters θ, here Easting

θ = [C0 C1 R]T
Fig. 2. Sample sites, each circle is centered on a sample point, radius is pro-
∗ 2 portional to the quantity measured, in this case the height above the ground
min kγ̂(h) − γ (θ, h)k . water in a 10 km × 10 km area near Slagelse, Denmark.
θ
 ALLAN A. NIELSEN: GEOSTATISTICS AND ANALYSIS OF SPATIAL DATA 3

ha a Gauss an mode may perform be er han he exponen a

mode n h s case Therefore a Gauss an mode based on he
number of data 333
exper men a sem -var ogram s shown a so The nugge e ec
mean 26.9839
.100 std.dev. 19.3422 s 18 m2 he range s 1 890 m and he s s 364 m2
coeff. of var. 0.7168
maximum
75% fractile
84.0000
40.0000
III S PAT AL I NTERPOLAT ON
.080 median 27.0000
25% fractile 7.5000 Th s sec on dea s w h de erm n s c ypes of n erpo a on
minimum 1.0776
such as nverse d s ance we gh ng and rad a bas s func ons
Frequency

.060 and s a s ca ypes known under he o n name of kr g ng

Spec fica y equa ons for s mp e and ord nary kr g ng are de-
.040 duced
O her de erm n s c n erpo a on me hods use (De aunay) r -
angu a on Vorono esse a on regress on ana ys s for de erm -
.020
na on of rend surfaces m n mum curva ure e c [15] [3]

.000 A Inverse D s ance We gh ng

0. 20. 40. 60. 80. 100. Poss b y he s mp es conce vab e way of carry ng ou n er-
Height po a on cons s s of ass gn ng he va ue of he neares ne ghbour
o a po n where he va ue s unknown An po en a mprove-
Fig. 3. Simple statistics and histogram for height data near Slagelse, Denmark.
men cons s s of ass gn ng h gher we gh s o observa ons c oser
o he po n s o wh ch we n erpo a e An obv ous way of do-
ng h s s o ass gn we gh s ha are propor ona o he nverse
d s ance from he des red po n o a N po n s en er ng n o he
n erpo a on For he i h po n we ge he we gh

1/d
w = PN
. .. .. .. .... ...... . . ... . ...
..... .......... ... .
. j=1 1/dj
.. .... .
6000

. .
. .. . .... .. ....... . . .
. ....... ........ .......... . ... . . ................ ..............
. . . . . .... . .. . .
.
.... . .............. .......... .. ..... .
. where dj s he d s ance from po n j o he po n o wh ch we
.. . ........... . ............. . ............. .. .. ... ...................................................... ........... . . n erpo a e Th s s read y ex ended o we gh ng w h d fferen
.. ........ . ......... ..... ............... .. . . . ... ............................................................... .
4000

. .. ... ... .. .. . ................. ..... .......... ... .. . .................. ..................................................................... .. powers p > 0 of he nverse d s ance
.. . .... . ............... .... .... . ................ ... .......... .... . . ........ . .
.. . . .... ...... ........... ...... .... .......................... ....... .... ..... ............ ................ .. .. .
. . .. ........... . ..... ............................................................................................................................................... .......................... ....
. . .. ...... ... ........ ......................... .......................................... .......... .......................................................... .......... ......... ...... ... ........ ... ..
. ....... ........ .......................... ..................................................................................................................................................
.................................................................................................................
.. .. .... ....... ..... ....................... .
...................................................................................................................................................................................................... . 1/dp
.......
.. w = PN
2000

. . . . . .. .. .. ... . . . .. . . .
.................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................. ... ..
. . . .. . .. . . . . . . . . . . . . . .
p
j=1 1/dj
. . ....................................... ...........
. ...................................................................... ............................................................... ........................................................................................................................................................................................................... .
.......................................
... . . ............... ........... .............................................................
..............
............................................................................
............................................................................... .......................................................................................................................................... .
........................................................................................................................... . . . . . . .........
. ................................
.................... ................................................................................................................................................................................. ............. .. . . ....
.
...... ...... ..................................................................................................................................................... . . . .. . . . . . . .
.. . .
....................... . . .
............
..
..
.....
..
.......
. . .. .
. . .. ..
. .
......... .
...
. . .
.....
. . . .
. . .
.. .
............
. . .... .............................................. .................................................................................................................................................................................................................................................................................................................. ...
..... .......... ............... . ....
.... .........................................................................................
......... .. ................
.. .......................
. ..
... .........................
. ..
. .......
..
.. . . ...
. ..
. ...
.
.. ...
.
................................
.
..... . .... . ... . .
.............................................................................................
. . . . . . . ..
. . . . . . .........
......
. ...................................................................................................................................
. ................................................................................................................................................................................... ............ . .
..
..........
............
.................
.....
...
....
......
............................
..................
..........
.....
.. .. ......
.......
..
... .
..........
.........
...... .........
...
... ...................................
.............
....
.................
.......... ......
..
... . .
.. .
.
................................. .
.
.
.
........
.
..
. ............
.
...
.................................
...
.
. ....
. ... . . . ...
.
.
............................
.....
.
.
......
.
.... .
. .
..............................
.........
.
..........
...... ........
.
.......................
............. .....
.
....
.
....................................................................................
.
............
...... ...... .....
...........
.......
.
.
...... .
.
.....
.
... ....
. .....
....
..
..........................
...
................
......
.................. .......... .
.
....
.
.........
...
.
..
.
......
...
. .
. .
.
.........
............
.
. .
. .
.
............
.....
..
.
.
..........
.
..
. .
...
. .
.
.............
......
...... ...
.
. .
. . . ....
......................
...
.
.. .........
... .
... .
...
. .
.
.........
...................................
.
..
. .
..
. .. ........
....................
..........
.....
............
........................
. .........
......................
....... ................
............ ............. ...
...
.....
.................
...
.............................................................................................
........................................
..
... ................................. .................
............................................................................................................................... ... .. .
.............
..... ......
.....
... .....
.... ...... ......
.............. ....
.... ....... ...........
......... ...................
..... ...........
............................. ..... . ..
A 1 Examp es
0

0 2000 4000 6000 8000 10000 12000 We now w sh o n erpo a e o Z0 a pos on r = 0 n F gure 1
d s ance m b m o nverse d s ance we gh ng d s he d s ance from po n
Z o Z0 We read y ca cu a e he fo ow ng we gh s
F g 4 A poss b e pa w se squa ed d e ences as a unc on o he magn ude
o he d sp acemen vec o exponen a va og am mode shown P
r d 1/d (1/d )/ (1/d )
−2 2 1/2 3/11 (= 0 2727)
−1 1 1 6/11 (= 0 5455)
3 3 1/3 2/11 (= 0 1818)

For d fferen powers of d we ge he we gh s

r d p=01 p = 2 0 p = 10 0
500

−2 2 0 3298 0 1837 0 0010

400

−1 1 0 3535 0 7347 0 9990

3 3 0 3167 0 0816 0 0000
300

We see ha for ow va ues of p he we gh s approach 1/N for

200

a po n s used For h gh va ues of p we ge near a we gh of one

for he neares ne ghbour
100

B Rad a Bas s Func on In erpo a on

0

0 2000 4000 6000 Cons der a near es ma e ẑ0 = ẑ(r 0 ) a oca on r 0 based
d s ance m
on N measuremen s z = z(r 1 ) z(r N ) T = z1 zN T
T
F g 5 Expe men a sem va og am as a unc on o he magn ude o he a oca ons r 1 rN Assume ha each observa on nflu-
d sp acemen vec o Gauss an sem va og am mode shown ences s surround ngs n he same way n a d rec ons and ha
he nfluence s expressed by some func on φ ( he rad a bas s
func on RBF) wh ch depends on he d s ance h = kr 0 − r k
4 ALLAN A. NIELSEN: GEOSTATISTICS AND ANALYSIS OF SPATIAL DATA

between locations r 0 and r i only, φ = φ(kr 0 − r i k). We shall B.2 Choice of RBF
look into choices of φ shortly. Define the interpolated value
Often one of the following choices is made for φ
N
X • multiquadric: φ(h) = (h2 + h20 )1/2 ,
2 2 −1/2
ẑ0 = wi φ(kr 0 − r i k), • inverse multiquadric: φ(h) = (h + h0 ) ,
2
i=1 • thin-plate spline: φ(h) = h log(h/h0 ) (which tends to 0 for
h tending to 0), or
where wi is the weight associated with location i. Let us deter- 1
• Gaussian: φ(h) = exp(− 2 (h/h0 ) ),
2
mine the wi so that the interpolation becomes exact at the known
where h0 is a scale parameter to be chosen. Generally, h0 should
locations r j , i.e.,
be chosen larger than a typical distance between samples and
N smaller than the size of the study area. The multiquadric is said
X
zj = wi φ(kr j − r i k), j = 1, . . . , N. to be less sensitive to the choice of h0 . Especially the Gaussian
i=1 is sensitive to this choice.
With a Gaussian RBF you don’t need the polynomials men-
This makes up N equations with N unknowns, the wi , which tioned above, with the thin-plate spline RBF a linear polynomial
can be written in matrix form may be needed.
    
φ(kr 1 − r 1 k) · · · φ(kr 1 − r N k) w1 z1
 .. .. ..   ..   ..  B.3 Shepard Interpolation
 . . .  .  =  . 
A special case for the normalized RBF interpolation consists
φ(kr N − r 1 k) · · · φ(kr N − r N k) wN zN of setting the matrix on the left hand side to a constant times the
or identity matrix. This corresponds to applying a φ that tends to
infinity for h tending to 0, and is finite for h > 0. This leads
Φ w = z. to setting the weights equal to the measurements wi = zi , i.e.,
we needn’t solve the system of equations for w. In this case
Often a polynomial P (r j ) = c0 + c1 xj + c2 yj + · · · = φ(h) = h−p , 1 < p ≤ 3 with appropriate handling for h = 0 is
[1 xj yj . . .] c where c = [c0 c1 c2 . . .]T is the vector of coeffi- often used.
cients for the polynomial is added to the interpolation at location
(in two dimensions) r j = [xj yj ]T C. Kriging
N Kriging (after the South African mining engineer and profes-
X
zj = P (r j ) + wi φ(kr j − r i k), sor Danie Krige) is a name for a family of methods for minimum
i=1 error variance estimation. Consider a linear (or rather affine) es-
timate ẑ0 = ẑ(r 0 ) at location r 0 based on N measurements
so that P T w = 0. Here z = [z(r 1 ), . . . , z(r N )]T = [z1 , . . . , zN ]T
 
1 x1 y1 x21 y12 x1 y1 ··· N
X
 .. .. .. .. .. .. ..  ẑ0 = w0 + wi zi = w0 + wT z,
P =  . . . . . . . 
i=1
1 xN yN x2N 2
yN xN yN ···
where wi are the weights applied to zi and w0 is a constant.
defines the polynomial applied. Often a constant corresponding
We consider zi as realisations of stochastic variables Zi , Z =
to the column of ones or a linear polynomial is used.
[Z(r 1 ), . . . , Z(r N )]T = [Z1 , . . . , ZN ]T . We think of Z(r) as
Solving the system of equations Φ w = z under the con-
consisting of a mean value and a residual Z(r) = µ(r) + ²(r)
straint P T w = 0 we get (see above on ordinary kriging and the
with mean value zero and constant variance σ 2 , E{²} = 0 and
Lagrange multiplier technique)
Var{²} = σ 2 . For the linear estimator we get
· ¸· ¸ · ¸
Φ P w z
= . Ẑ0 = w0 + wT Z. (2)
PT 0 c 0

The order of the polynomial may depend on the RBF chosen. The estimation error z0 − ẑ0 is unknown. But for the expec-
References on RBF are [16], [17], [18]. tation value of the estimation error we get

B.1 Normalized RBF Interpolation E{Z0 − Ẑ0 } = E{Z0 − w0 − wT Z}

RBF interpolation comes in a normalized version also = µ0 − w0 − wT µ, (3)
PN
i=1 wi φ(kr 0 − r i k) where µ0 = µ(r 0 ) is the expectation value of Z0 and µ is a
ẑ0 = PN vector of expectation values for Z
i=1 φ(kr 0 − r i k)
   
which leads to µ(r 1 ) µ1
 ..   .. 
µ =  .  =  . .
N
X N
X
zj φ(kr j − r i k) = wi φ(kr j − r i k), j = 1, . . . , N. µ(r N ) µN
i=1 i=1
We want our estimator to be unbiased or central, i.e., we de-
for the weights. In matrix form this leads to the same matrix mand E{Z0 − Ẑ0 } = 0 or
equation as above with the elements in the vector on the right
PN
hand side changed from zj to zj i=1 φ(kr j − r i k). µ0 − w0 − wT µ = 0. (4)
ALLAN A. NIELSEN: GEOSTATISTICS AND ANALYSIS OF SPATIAL DATA 5

The variance of the estimation error is for any µ0 . 1 is a vector of ones. This is possible only if w0 = 0
and wT 1 = 1.
2
σE = Var{Z0 − Zˆ0 } The weights wi are found by minimising σE 2
under the con-
T
= Var{Z0 } + Var{w0 + wT Z} straint w 1 = 1. A standard technique for minimisation under
−2 Cov{Z0 , w0 + wT Z} under a constraint is introducing a function F with a so-called
Lagrange multiplier (here −2λ) which we multiply by the con-
= σ 2 + wT (Cw − 2 Cov{Z0 , Z}), straint set to zero and then minimising
where C is the dispersion or variance-covariance matrix of the F = 2
σE + 2 λ(wT 1 − 1)
stochastic variables, Z, entering into the estimation.
What is said in Section III-C so far is valid for all linear es- without constraints. Again the partial derivatives are set to zero
timators. The idea in krigingis now to find the linear estimator
which minimises the estimation variance. ∂F
= 2 Cw − 2 Cov{Z0 , Z} + 2 λ1 = 0
∂w
C.1 Simple Kriging ∂F
= 2 (wT 1 − 1) = 0,
In simple kriging (SK) we assume that µ(r) is known. From ∂λ
Equations 2 and 4 we get which results in the OK system
Ẑ0 − µ0 = wT (Z − µ). Cw + λ1 = Cov{Z0 , Z}
The weights wi are found by minimising the estimation variance 1T w = 1
2
σE . This is done by setting the partial derivatives to zero
or
2
∂σE     
= 2 Cw − 2 Cov{Z0 , Z} = 0, C11 ··· C1N 1 w1 C01
∂w  .. .. .. ..   ..   .. 
 . . . .     . 
  .  =  .
which results in the SK system  CN 1 ··· CN N 1   wN   C0N 
1 ··· 1 0 λ 1
Cw = Cov{Z0 , Z}
The values requested for Cij are found as described in the pre-
or vious section on SK.
     The minimised squared estimation error termed the ordinary
C11 ··· C1N w1 C01
 .. .. ..   ..  =  ..  kriging variance is
 . . .  .   . ,
2
CN 1 ··· CN N wN C0N σOK = σ 2 + wT (Cw − 2 Cov{Z0 , Z})
= σ 2 − wT Cov{Z0 , Z} − λ.
where Cij , i, j = 1, . . . , N is the covariance between points i
and j among the N points, which enter into the estimation of OK implies an implicit re-estimation of µ0 for each new con-
point 0. C0j , j = 1, . . . , N is the covariance between point j stellation of points. This is an attractive property making OK
and point 0, the point to which we interpolate. We get these co- well suited for interpolation in situations where the mean is not
variances from the semi-variogram model (remembering Equa- constant (i.e., in the absence of first order stationarity).
tion 1, γ(h) = C(0) − C(h)) as the sill minus the value of
the semi-variogram model for the relevant distance (and possi- C.3 Examples
bly direction) between observations. (Alternatively, the kriging Let us consider the data in Figure 1 again. We now wish to
system may be formulated b.m.o. the semi-variogram; to avoid interpolate to the position r = 0 b.m.o. ordinary kriging. To
zeros on the diagonal of C we prefer the covariance formulation carry out the calculations we use a stipulated semi-variogram
for numerical reasons.) Here Cij must not be confused with the based on the spherical model with C0 = 0, C1 = 1 and R = 6.
semi-variogram parameters C0 and C1 . Remembering Equation 1, C(h) = C(0) − γ(h), this gives the
The minimised squared estimation error termed the simple auto-covariance function (in this case where C0 + C1 = 1 this
kriging variance is is the same as the auto-correlation function)
2
σSK = σ 2 + wT (Cw − 2Cov{Z0 , Z}) h γ̂(h) C(h)
= σ 2 − wT Cov{Z0 , Z}. 0 0.0000 1.0000
1 0.2477 0.7523
In SK the mean value µ(r) is known. In practice it is often 2 0.4815 0.5185
assumed constant for the entire domain (or study area), or we 3 0.6875 0.3125
must estimate it before the interpolation (or we must construct 4 0.8519 0.1481
an interpolation algorithm which does not require knowledge of 5 0.9606 0.0394
the mean field, see the next section). 6 1.0000 0.0000
C.2 Ordinary Kriging
Therefore the OK system looks like this
In ordinary kriging (OK) we assume that the mean µ(r) is     
constant and equal to µ0 for Z0 and the N points that enter into 1.0000 0.7523 0.0394 1 w1 0.5185
the estimation of Z0 . From Equations 3 and 4 we get  0.7523 1.0000 0.1481 1   w2   0.7523 
  = 
 0.0394 0.1481 1.0000 1   w3   0.3125 
E{Z0 − Ẑ0 } = µ0 (1 − wT 1) − w0 = 0 1 1 1 0 λ 1
6 ALLAN A. NIELSEN: GEOSTATISTICS AND ANALYSIS OF SPATIAL DATA
80.0
72.0
64.0
56.0
48.0
40.0
32.0
24.0
16.0
8.0
.0
Fig. 6. Kriged map of heights above the ground water (unit is m).
400.
360.
320.
280.
240.
200.
160.
120.
80.0
40.0
.0
Fig. 7. Kriging variance corresponding to Figure 6 (unit is m2 ).
where the values for Cij come from the C(h) table. The so-
lution is w1 = −0.0407, w2 = 0.7955, w3 = 0.2452 and
λ = −0.0489, which gives a kriging variance of 0.3949. We
see that even though Z1 is closer to Z0 than Z3 , the weight on
Z1 is much smaller than the weight on Z3 . This is an attractive
characteristic of kriging, which allows for possible clustering of
the sampling locations. We say that Z2 screens for Z1 . This
screening effect becomes weaker for higher nugget effects and
it disappears for pure nugget effect (i.e., C1 = 0 for the models
shown here), where all weights become equal.
In a more realistic example Figure 6 shows a kriged (OK) map
over heights above the ground water for the area near Slagelse.
The interpolation is based on the isotropic Gaussian model for
the experimental semi-variogram in Figure 5 (nugget effect 18
m2 , range 1.890 m and sill 364 m2 ). We have kriged to 100 by
100 points in a 100 m by 100 m grid using a moving window
to include the points from which to interpolate. The search ra-
dius of the moving window was 2,000 m and the estimation for
each point was based on minimum 4 and maximum 20 points.
We see that the interpolated map shows a good correspondance
with the map of sample sites in Figure 2. Figure 7 shows the
corresponding OK variances. We see that the kriging variance
is large where the distances to the nearest samples are large.
C.4 Other Types of Kriging
If we wish to estimate average (also known as regularised)
values over an area or a volume rather than point values, we
may use block kriging which can be combined with several other
forms of kriging.
If more variables are studied simultaneuously the above meth-
ods for description of spatial correlation may be extended to
handle the spatial covariation between all pairs of variables in
the form of cross semi-variograms or cross-covariance func-
tions. Also more variables may be interpolated simultaneuously
using cokriging. Cokriging is most useful when one variable is
sampled on fewer locations than other correlated variables. Uni-
versal kriging is a method for the case where the mean value is
described by linear combinations of known functions ideally de-
termined by the physics of the problem at hand. Also methods
for non-linear kriging exist such as lognormal kriging, multi-
Gaussian kriging, rank kriging, indicator kriging and disjunc-
tive kriging. References here are [1], [3], [9], [14].
C.5 Conclusions, kriging
The above sections and examples demonstrate the following
properties of kriging:
• Kriging is a type of interpolation that gives us both an estimate
based on the spatial structure of the variable in question as ex-
pressed by the auto-covariance function (or the semi-variogram)
as well as an estimation variance which is minimised.
• The kriging estimator is the best linear unbiased estimator
(BLUE) in the sense that it minimises the estimation variance.
Also it is exact, i.e., if we interpolate to a point which coincides
with an existing sample point, kriging gives the same value as
the one measured and the kriging variance is zero.
• The kriging system and the kriging variance depend on the
auto-covariance function (or the semi-variogram) and the spatial
layout of the sample locations only and not on the actual data
values. If an auto-covariance function (or a semi-variogram)
is known (or assumed or imposed) this has important potential
for minimising the estimation variance in experimental design
(i.e., in the planning phase of the spatial lay-out of the sampling
scheme).
• The solution of the kriging system implies a statistical dis-
tance weighting of the data values which enter into the inter-
polation. Also for OK, the weights are scaled so they add to
one. Furthermore, possible redundancy in the form of clustering
of the sample locations is accounted for; the above mentioned
screening effect is due to this de-clustering.
• Because of the implicit re-estimation of the mean value for
each new point constellation, OK is well suited for situations
where the mean is not constant over the study region, i.e., where
we don’t have first order stationarity.
Further, the kriging system has a unique solution if and only if
the covariance matrix C (Section III-C) is positive definite; this
also guarantees a non-negative kriging variance.
The strength of kriging may be attributted to a combination
of the above characteristics.
If we choose to formulate the kriging system in terms of the
auto-covariance function which is done in this note, we must as-
sume (or impose) second order stationarity, i.e., the same auto-
covariance function over the entire study area. The system may
also be formulated in terms of the semi-variogram and in this
case we must assume the intrinsic hypothesis, i.e., the same
semi-variogram over the entire study area.
These assumptions may seem to be a drawback of kriging but
if deterministic methods are applied, we implicitly make similar
ALLAN A. NIELSEN: GEOSTATISTICS AND ANALYSIS OF SPATIAL DATA 7

assumptions. It can hardly be considered as a drawback of geo- [19] E. C. Grunsky and F. P. Agterberg, “Spatial and multivariate analysis of
statistical methods that we are forced to consider whether such geochemical data from metavolcanic rocks in the Ben Nevis area, On-
tario,” Mathematical Geology, vol. 20, no. 7, pp. 825–861, 1988.
assumptions are appropriate. [20] E. C. Grunsky and F. P. Agterberg, “SPFAC: a Fortran program for spatial
factor analysis of multivariate data,” Computers & Geosciences, vol. 17,
IV. F INAL R EMARKS no. 1, pp. 133–160, 1991.
[21] Allan A. Nielsen, Knut Conradsen, John L. Pedersen, and Agnete Steen-
A sampling strategy may be based on the dependence of the felt, “Spatial factor analysis of stream sediment geochemistry data from
South Greenland,” in Proceedings of the Third Annual Conference of the
kriging variance on the distance to the nearest samples. If the International Association for Mathematical Geology, Vera Pawlowsky-
auto-covariance function (or the semi-variogram) and the sam- Glahn, Ed., Barcelona, Spain, September 1997, pp. 955–960.
ple locations are known, we can determine the kriging weights [22] Allan A. Nielsen, Knut Conradsen, John L. Pedersen, and Agnete Steen-
felt, “Maximum autocorrelation factorial kriging,” in Proceedings of the
and the kriging variances before the actual sampling takes place. 6th International Geostatistics Congress (Geostats 2000), W. J. Kleingeld
If the variances become too large in some regions of our study and D. G. Krige, Eds., Cape Town, South Africa, April 2000. Internet
area we may modify the sample locations to obtain smaller vari- http://www.imm.dtu.dk/pubdb/p.php?3639.
[23] Knut Conradsen, Introduktion til Statistik, Informatics and Mathematical
ances. Also, to obtain a good estimate of the nugget effect which Modelling, Technical University of Denmark, 1984.
is an important parameter for the outcome of the kriging process, [24] T. W. Anderson, An Introduction to Multivariate Statistical Analysis, sec-
it may be an advantage to position some samples close to each ond edition, John Wiley, New York, 1984.
[25] D. A. Griffith and C. G. Amrhein, Multivariate Statistical Analysis for
other. Geographers, Prentice Hall, 1997.
In multivariate studies where the joint behaviour of several [26] Annette Kjær Ersbøll, “A comparison of two spatio-temporal semivari-
variables is investigated, rather than interpolating the original ograms with use in agriculture,” in Lecture Notes in Statistics, T. G. Gre-
goire et al., Ed., vol. 122, pp. 299–308. Springer-Verlag, 1997.
variables we may interpolate combinations of them. For in- [27] Annette Kjær Ersbøll and Bjarne Kjær Ersbøll, “On spatio-temporal krig-
stance we may interpolate principal components or factors re- ing,” in Proceedings of the Third Annual Conference of the International
Association for Mathematical Geology (IAMG’97), Vera Pawlowsky-
sulting from a factor analysis or a spatial factor analysis, [19], Glahn, Ed., Barcelona, Spain, September 1997, pp. 617–622.
[20], [11], [21], [22]. Generel references to multivariate statis- [28] Thomas Knudsen, Busstop - A Spatio-Temporal Information System,
tics are for example [23], [24]. [25] is written especially for Ph.D. thesis, Niels Bohr Institute, Department of Geophysics, University
of Copenhagen, Denmark, 1997.
geographers, [15] for geologists.
Also temporal aspects in connection with the application of I NDEX T ERMS
data that vary in both space and time may be important. Spatio-
auto-correlation function
temporal semi-variograms and spatio-temporal kriging are dealt auto-covariance function
with in for example [26], [27]. A GIS for handling of temporal BLUE
data is described in [28]. data analysis
distance weighting
estimation variance
R EFERENCES geostatistics
[1] A. G. Journel and Ch. J. Huijbregts, Mining Geostatistics, Academic GIS
Press, London, 1978, 600 pp. GSLIB
[2] Isobel Clark, Practical Geostatistics, Elsevier Applied Science, London, intrinsic hypothesis
1979, http://www.stokos.demon.co.uk/. covariance function
[3] Edward H. Isaaks and R. Mohan Srivastava, An Introduction to Applied kriging
Geostatistics, Oxford University Press, New York, 1989, 561 pp. kriging variance
[4] K. Conradsen, A. A. Nielsen, and K. Windfeld, “Analysis of geochemical Lagrange multiplier
data sampled on a regional scale,” in Statistics in the Environmental and least squares method
Earth Sciences, A. Walden and P. Guttorp, Eds., pp. 283–300. Griffin, multivariate statistics
1992. nested structures
[5] Noel A. C. Cressie, Statistics for Spatial Data, revised edition, Wiley, normalized radial basis functions
New York, 1993. nugget effect
[6] P. A. Burrough and R. A. MacDonnell, Principles of Geographical Infor- ordinary kriging
mation Systems, Oxford University Press, 1998. sampling strategy
[7] M. F. Goodchild, B. O. Parks, and L. T. Steyaert, Environmental Modeling radial basis functions
with GIS, Oxford University Press, 1993. range of influence
[8] H. Wackernagel, Multivariate Geostatistics, third edition, Springer, 2003. regionalised variable
[9] C. V. Deutsch and A. G. Journel, GSLIB: Geostatistical Software Library screening effect
and User’s Guide, second edition, Oxford University Press, 1998, Internet semi-variogram
http://www.gslib.com/.
semi-variogram model
[10] Y. Pannatier, VARIOWIN: Software for Spatial Data Analysis in 2D,
sill
Springer, 1996, Internet http://www-sst.unil.ch/research/variowin/.
simple kriging
[11] Allan A. Nielsen, Analysis of Regularly and Irregularly Sampled Spa-
tial, Multivariate, and Multi-temporal Data, Ph.D. thesis, Informatics spatial interpolation
and Mathematical Modelling, Technical University of Denmark, Lyngby, spatial correlation
1994, Internet http://www.imm.dtu.dk/pubdb/p.php?296. stationarity
[12] Karsten Hartelius, Analysis of Irregularly Distributed Points, Ph.D. thesis, statistics
Informatics and Mathematical Modelling, Technical University of Den- stochastic function
mark, Lyngby, 1996. Internet http://www.imm.dtu.dk/pubdb/p.php?1204. stochastic process
[13] Allan A. Nielsen, “2D semivariograms,” in Proceedings of the Fourth stochastic variable
South African Workshop on Pattern Recognition, Paul Cilliers, Ed., Si- variogram
mon’s Town, South Africa, 25–26 November 1993, pp. 25–35. Variowin
[14] Allan A. Nielsen, “Kriging,” lecture note, Technical University of Den-
mark, 2004. Internet http://www.imm.dtu.dk/pubdb/p.php?3479.
[15] J. C. Davis, Statistics and Data Analysis in Geology, second edition, John
Wiley & Sons, 1986.
[16] M. J. D. Powell, “The theory of radial basis function approximation,” in
Advances in Numerical Analysis II: Wavelets, Subdivision Algorithms and
Radial Functions, W. A. Light, Ed., Oxford University Press, 1992.
[17] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numer-
ical Recipes: The Art of Scientific Computing, third edition, Cambridge
University Press, 2007.
[18] F. Anton, J. A. Bærentzen, J. Gravesen, and H. Aanæs, Computational Ge-
ometry Processing, lecture note, Technical University of Denmark, 2008.